CN113313147A - Image matching method based on deep semantic alignment network model - Google Patents
Image matching method based on deep semantic alignment network model Download PDFInfo
- Publication number
- CN113313147A CN113313147A CN202110516741.5A CN202110516741A CN113313147A CN 113313147 A CN113313147 A CN 113313147A CN 202110516741 A CN202110516741 A CN 202110516741A CN 113313147 A CN113313147 A CN 113313147A
- Authority
- CN
- China
- Prior art keywords
- image
- transformation
- alignment
- model
- regression
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 67
- 230000009466 transformation Effects 0.000 claims abstract description 133
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims abstract description 61
- 230000002457 bidirectional effect Effects 0.000 claims abstract description 18
- 230000000694 effects Effects 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims abstract description 10
- 238000013519 translation Methods 0.000 claims abstract description 3
- 238000012549 training Methods 0.000 claims description 30
- 238000013527 convolutional neural network Methods 0.000 claims description 26
- 238000001514 detection method Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 10
- 238000000844 transformation Methods 0.000 claims description 10
- 238000013459 approach Methods 0.000 claims description 9
- 238000004458 analytical method Methods 0.000 claims description 7
- 230000005540 biological transmission Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000013461 design Methods 0.000 claims description 5
- 238000012546 transfer Methods 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 239000013598 vector Substances 0.000 claims description 3
- 238000009825 accumulation Methods 0.000 claims description 2
- 238000011156 evaluation Methods 0.000 claims description 2
- 230000008447 perception Effects 0.000 claims description 2
- 239000004576 sand Substances 0.000 claims description 2
- 238000002372 labelling Methods 0.000 claims 1
- 238000013507 mapping Methods 0.000 claims 1
- 230000011218 segmentation Effects 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000001131 transforming effect Effects 0.000 description 3
- 241000282465 Canis Species 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 240000008313 Pseudognaphalium affine Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 239000002904 solvent Substances 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Multimedia (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image matching method based on a deep semantic alignment network model, which comprises the steps of gradually estimating the alignment between two semantic similar images by establishing an object position-aware semantic alignment network model OLASA; the method is characterized in that a triple sampling strategy is adopted to train a network model OLASA, and three sub-networks N of potential object co-location POCL, affine transformation regression ATR and bidirectional thin plate spline regression TTPS are adoptedtran,NaffiAnd NttpsRespectively estimating translation, affine transformation and spline transformation; and then, establishing and optimizing an alignment relation between the images in a layering manner to obtain an image matching result. By utilizing the technical scheme provided by the invention, the image alignment effect with larger position difference can be improved,the accuracy of image matching is improved. The method can be applied to target tracking, semantic segmentation, multi-view three-dimensional reconstruction and the like in the field of computer vision.
Description
Technical Field
The invention belongs to the technical field of computer vision and digital image processing, relates to an image matching technology, and particularly relates to a method for establishing an accurate corresponding matching relation of main target objects in similar images based on an image depth semantic alignment network model.
Background
Image semantic alignment aims at establishing accurate corresponding relation of similar target objects among images, namely point-to-point characteristic matching relation among the similar target objects in different images. The specific scene is to analyze and quantify the similarity between features by using the feature information of the images on the premise that the image content information is the same or similar, and further determine the matching relationship of feature points on similar objects in the images. The problem is a basic problem in computer vision, and the method is widely applied to the fields of target tracking, image semantic segmentation, multi-view three-dimensional reconstruction and the like.
Semantic alignment has received much attention in recent years. Early studies included methods of finding instance-level matches by defining and computing sparse or dense descriptors [5 ]. However, the example-level descriptions of these methods lack the generalization capability corresponding to the category level. The purpose of category-level correspondence is to find dense correspondence between semantically similar images. Some methods use local descriptors and minimize the matching energy invested. And the manually constructed descriptor is difficult to embed high-level semantic features and sensitive to image change.
Inspired by the rich high-level semantics of Convolutional Neural Network (CNN) features, recent solutions (references [2], [4], [6], [7], [8]) employ training CNN features and combine them to estimate dense flow fields, aligning the images. Furthermore, the methods employed in references [1], [3], [9], [10] estimate the geometric transformation with trainable CNN features and express semantic correspondences as geometric alignment problems. Some of these methods outperform dense flow based methods and produce smoother matching results, thanks to the geometric transformations describing dense correspondences.
Despite the great advances made by existing approaches, the semantic alignment problem still faces challenges such as alignment difficulties caused by object variations (e.g., appearance, scale, shape, position) and complex backgrounds. Specifically, firstly, since the target position difference is large, it is difficult to directly establish a dense correspondence relationship between images, and the effect is not good (as shown in fig. 1). Previous methods often fail to align such images due to insufficient research to handle such situations. Secondly, it is a difficulty in data annotation that it is difficult to collect a large number of training image pairs with ground truth dense correspondences and significant appearance transformations. Manually annotating such training data is very labor intensive and is somewhat subjective.
Reference documents:
[1]Ignacio Rocco,Relja Arandjelovic,and Josef Sivic,“Convolutional neural network architecture for geometric matching,”inCVPR,2017.
[2]Kai Han,Rafael S Rezende,Bumsub Ham,Kwan-Yee K Wong,Minsu Cho,Cordelia Schmid,and Jean Ponce,“Scnet:Learning semantic correspondence,”in ICCV,2017.
[3]Ignacio Rocco,Reljaand Josef Sivic,“End-to-end weakly-supervised semantic alignment,”in CVPR,2018.
[4]Junghyup Lee,Dohyung Kim,Jean Ponce,and Bumsub Ham,“Sfnet:Learning object-aware semantic correspondence,”in CVPR,2019.
[5]David G Lowe,“Distinctive image features from scale-invariant keypoints,”IJCV,vol.60,no.2,2004.
[6]Ce Liu,Jenny Yuen,and Antonio Torralba,“Sift flow:Dense correspondence across scenes and its applications,”IEEE TPAMI,vol.33,no.5,2010.
[7]Bumsub Ham,Minsu Cho,Cordelia Schmid,and Jean Ponce,“Proposal flow:Semantic correspondences from object proposals,”IEEETPAMI,vol.40,no.7,2017.
[8]Seungryong Kim,Dongbo Min,Bumsub Ham,Sangryul Jeon,Stephen Lin,and Kwanghoon Sohn,“Fcss:Fully convolutional self-similarity for dense semantic correspondence,”in CVPR,2017.
[9]Paul Hongsuck Seo,Jongmin Lee,Deunsol Jung,Bohyung Han,and Minsu Cho,“Attentive semantic alignment with offset-aware correlation kernels,”in ECCV,2018.
disclosure of Invention
In order to overcome the defects of the prior art, the invention provides an image matching method based on a deep semantic alignment network model, which establishes a semantic alignment network for sensing the position of an object and trains the network by adopting a triple sampling strategy to establish and optimize the alignment relationship between images in a layered manner, solves the technical problems that the prior art is difficult to directly establish the dense corresponding relationship between the images and the image data annotation wastes time and labor and has low accuracy, and improves the accuracy of image matching.
The image semantic alignment technology in the invention is a sub-problem in the field of image matching or image feature matching, and mainly aims at the following scenes: although the two images to be matched are different, the two images to be matched both contain a similar foreground object, namely, the high-level semantic information such as the appearance, the shape, the posture and the like of the foreground object are similar, and the object basically belongs to the same category, such as cars of different brands.
The technical scheme provided by the invention is as follows:
an image matching method based on a deep semantic alignment network model comprises the steps of gradually and robustly estimating alignment between two semantic similar images by establishing an object position-aware semantic alignment network model (OLASA); meanwhile, a triple sampling strategy is provided for training the network, translation, affine transformation and spline transformation are respectively estimated through three sub-networks (potential object co-location (POCL), Affine Transformation Regression (ATR) and bidirectional thin plate spline regression (TTPS)), and then the alignment relation between images is established and optimized in a layering mode to obtain an image matching result; the method comprises the following steps:
step 1, extracting semantic features of an image;
in the method, an independent Convolutional Neural Network (CNN) is adopted at the front end of each sub-network to extract the characteristics of the image. In specific implementation, the invention adopts a Convolutional Neural Network (CNN) to extract the characteristics of two images, and the network can be the most basic CNN network or the CNN network after improvement or enhancement.
In the formula (1), F is a feature extracted from an image,the method comprises the following steps that (1) a characteristic data space (real number) is formed, and h, w and d respectively represent three dimensions of the characteristic data space, namely height, width and channel number;is a Convolutional Neural Network (CNN); i is an image;h, W, D represent three dimensions of the image data space, i.e. height, width, number of channels, respectively, for the data space (real number) of the image.
A pair of imagesAsRespectively extracting two groups of obtained features Is,ItRespectively a source image and a target image; fs,FtRespectively a source image semantic feature and a target image semantic feature.
The invention establishes an object position-aware semantic alignment network model-OLASA, and gradually and robustlyThe alignment between two semantically similar images is estimated. The system architecture of OLASA also takes POCL, ATR, and TTPS as main components, and they are named N respectivelytran,Naffi and NttpsFor estimating offset, affine and TTPS transformations, respectively, denoted Ttran,Taffi and TttpsAs shown in fig. 3. By transforming the model Ttran,Taffi and TttpsThe transformation result of the source image at each stage can be obtainedAndcontinuous transformation I using these transformation modelsHI.e. byObtaining a source image IsWith the target image ItAlignment result ofThe method for establishing the semantic alignment network model for object position perception comprises the steps 2-4.
Step 2, adopting potential objects to cooperatively locate the sub-network (N)tran) Estimating the offset of a target object between images, and eliminating the position deviation of an object to be matched;
aiming at the problem that similar target objects among images often have obvious displacement, NtranThe sub-network adopts the potential target position detection and estimation technology to predict an offset transformation model in advance and eliminate the influence of the source image on a large span by transforming the source image.
The potential object co-location sub-network is denoted asFeature F of source imagesAnd features F of the target imagetAs input, estimating potential targets by a classified target detection method, and training a co-location sub-network according to the positions of the potential targetsAnd then realize Is and ItEstimation of preliminary transformation betweenIs represented as follows:
in the formula ,co-locating a sub-network for the potential object;is Is and ItAn estimate of the preliminary transform (offset transform) between;the degree of freedom of (a) is 4; (x, y) represents the spatial coordinates of the target image and (x ', y') represents the corresponding sample coordinates in the source image.
To estimate an offset transformation modelThe position of an object to be matched needs to be estimated firstly, and therefore, the approximate position of the object to be matched can be estimated by adopting the existing classified target detection technology;
further location Is and ItTwo major and semantically related objects; and then another feature extraction module, such as CNN,to extract the corresponding feature descriptor Vs i} and {Vt jI and j respectively represent the serial numbers of the feature points in the source image and the target image;
will describe the son { Vs i} and {Vt jStack (stack) as a feature matrixAnd the semantic similarity matrix is calculated by multiplying the two feature matrices,
selection of ZstForming a similar feature pair group by a plurality of feature pairs with the highest similarity scores, calculating two corresponding regions by using feature point groups respectively represented in the source image and the target image as two main potential objects, and calculating corresponding frame coordinates, namely space positions;
finally, obtaining the space coordinates of the frames corresponding to the two main potential objects by positioning, and calculating a position deviation transformation modelSource image IsBy passingTransforming into position-shifted images
Step 3, constructing affine transformation regression sub-network NaffiEstimating affine transformation model of image to be matched using affine transformation regression subnetworkObtaining affine transformation parameter estimation;
the ATR subnetwork is used to estimate an affine transformation model of the offset adjusted image and the target image. The ATR subnetwork needs to construct pairs of image features, i.e. feature pairs, and calculate the correlation of the feature pairs, from which the parameters of the affine transformation model are estimated.
In particular, the image subjected to position shift conversion isAnd a target image ItCharacteristic F ofs 1 and FtForming feature pairs, calculating the relevance of these feature pairsI.e. the 4D correlation tensorEach element of the tensorRecording two local feature vectorsThe inner product of (d) and (d). The eigenvectors and the correlation tensors are L2 normalized. Tensor of degree of correlationInput deviceTo carry out and ItInter affine transformation model estimation
Wherein the affine transformation modelIs 6, i.e. 6 affine transformation parameters need to be estimated. Image of a personCan be further transformed by the modelObtained after affine transformationAnd ItFurther alignment is accomplished.
Step 4, constructing a bidirectional thin plate spline regression subnetwork NttpsOptimizing the alignment effect by using a bidirectional thin plate spline regression subnetwork;
TTPS subnetwork NttpsAnd estimating a regression transformation model between the affine transformed image and the target image by using the control points, wherein the model can further improve or enhance the semantic alignment effect of the image. TTPS employs a bi-directional strategy to avoid excessive distortion or distortion of the image. Compared with the prior thin plate spline regression TPS method, the method increases the secondary image ItTo the imageAnd the control point adjustment in the opposite direction effectively removes excessive deformation and improves the matching distortion condition.
When embodied, in the image and ItA control point grid and a TTPS sub-network are arranged on the network for imaging and ItIs characterized by and FtThe correlation tensor of the feature pair is used as a calculation pairImage, i.e. 4D correlation tensorUsing TTPS regression sub-networksEstimating a regression deformation modelThe process of (d) can be expressed as:
wherein ,according to and FtCalculating the correlation of the middle control point pair to obtain a regression deformation modelIs obtained by calculating 6 transform weight parameters of the two. According to the estimatedRegression transformation model, canIs further transformed intoThereby realizing the source image and the target image ItIs aligned.
Step 5, jointly training an integral OLASA network model comprising three sub-networks;
in terms of the selection of training samples, the invention proposes a triplet method by introducing reference images, which can better capture geometrical and appearance variations in the training data.
Step 51, the present invention proposes a triple sampling strategy to generate training data. Each triplet containing a source image IsReference picture IrAnd a target image It. Wherein, from IsBy random geometric transformationOrGenerated reference image Ir。
And step 52, designing three loss functions to realize the optimization of the OLASA, wherein the three loss functions comprise transitivity loss, consistency loss and alignment loss.
Transfer loss functionFor measuring the accuracy of the predicted transform model. To check whether the predicted transformation model is accurate, the transformation result of the model may be compared with the transformation result of another transformation approach for achieving the same transformation objective; another transformation approach may specifically employ an indirect transformation approach through a combination of two transformation models; the indirect transformation approach achieves the same transformation goal as the direct transformation approach which only adopts one transformation model, namely the transformation goal of the transformation model to be detected is the same.
To check a certain affine transformationFor example, if one is to be used(affine transformation from source image to reference image) as predicted transformation model, and checking its accuracy to find out another indirect transformation approach capable of implementing uniform transformation, such as combination of two transformation modelsI.e. by combination(affine transformation from source image to target image) and(affine transformation from target image to reference image) two related successive transformations to simulateTransformation, judging the transformation from the source image to the reference image by comparing the difference of the transformation results obtained in the two waysTo the degree of accuracy of the measurement. The specific implementation can be realized by means of a grid consisting of selected pixel points in the image, for example, twenty parts of the image are divided in the horizontal and vertical directions, so that a grid G (| G | ═ 20 × 20) can be obtained, wherein 400 points are included in the grid G, one point in G is represented by (x, y), and the initial grid can be constructed in IsThe above. As described hereinbefore, byAndto judgeThe accuracy implementation may be shifted to a specific calculation based on grid points, i.e.,theoretically, byShould be equal toThe projected points after successive deformations are aligned while in contrast with the real data, i.e.(affine transformation of source image to target image calculated from annotation data, not predicted)The difference is lost as an error on which the machine learning method is based, and therefore, the transfer loss functionCan be expressed as:
the transformation involves three types of objects, namely a source image, a target image and a reference image, and therefore the three types of objects are constructed into a triple for training a relevant model, namely a piece of training data is composed of a source image, a target image and a reference image. The method can also be seen as adding a corresponding reference image on the basis of the traditional training pair (source image and target image). The associated loss calculation makes it possible to simultaneously calculate three transformations, each involving a certain pair of images of the triplet (the two images belonging to different image types), e.g. affine transformation regression subnetworkEstimating ternary samples(source image and target image) and(target image and reference image) affine transformation model.
Similarly, transmission losses will also occur in this processFor spline regressionThe difference analysis of (3).
Consistency loss functionLoss of consistencyIs defined as: consistency loss functionIs defined as: in a ternary sample consisting of a source image, a target image and a reference image, the accumulation of the bidirectional reprojection error of a point in the grid G between different image pairs constitutes a consistency loss functionFor representing affine transformationsSum spline regressionA difference of (a);
given an image Is and ItInter affine transformationAnd corresponding inverse transformationWhen compared to the results of its two-way transformation, the point (x, y) e G should be aligned with the coordinates of the original space, i.e.,thus, the reprojection error of a point in grid G between a source image-target image pair can be calculated by the following equation:
wherein ,ε(Is,It) Is the reprojection error between the source-target image pair.
In the ternary sample composed of the source image, the target image and the reference image, other image pairs also contain bidirectional reprojection errors, and the two image pairs are all accumulated to be used as the consistency loss function of the method
Similarly, consistency is also lost in the present methodFor spline regressionThe difference analysis of (3).
Alignment loss functionBased on the literature [3]Soft-inline Count (Soft-inline Count) and two-way measurement alignment quality design for evaluationAndcan further optimize the alignment accuracy.The calculation formula is as follows:
wherein cstBy aligning IsTo ItSoft inline counting of, similarly, ctsIndicating alignment ItTo IsSoft inline counting of (2).
Step 6, obtaining an image matching result;
according to the joint training integral network, aiming at a given image pair to be matched, an alignment result from a source image to a target image can be given.
Specifically, learning training is carried out through an overall network comprising three sub-networks, and the method can obtain an image alignment transformation model according to a training data sample, so that a trained object position-aware semantic alignment network model OLASA is obtained. In the testing stage, the matching image result from the source image to the target image can be calculated for the given image pair to be matched.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides an image matching method based on a depth semantic alignment network model, which comprises three sub-networks of potential object co-location (POCL), Affine Transformation Regression (ATR) and bidirectional thin plate spline regression (TTPS), wherein the POCL can effectively sense the offset of a potential object, the ATR can learn the parameters of geometric deformation, the TTPS can improve the robustness of deformation, and the combined learning of the three sub-networks not only can realize the semantic alignment of images, but also can obtain the image matching effect with higher accuracy. By utilizing the technical scheme provided by the invention, the image alignment effect with larger position difference can be improved. Meanwhile, in a scene with insufficient annotation data, by means of generating a reference image, geometric changes and appearance changes in the existing data can be deeply mined and utilized, and the accuracy of image matching is improved. The method can be applied to various tasks in the field of computer vision, such as target tracking, semantic segmentation, multi-view three-dimensional reconstruction and the like.
Drawings
FIG. 1 is a diagram of object position-aware semantic alignment in images in image matching;
wherein Is represents the source image; it represents a target image.
FIG. 2 is a flow chart of a method for performing image matching on an OLASA network model established by the present invention;
FIG. 3 is a structural block diagram of an OLASA network model 2-2 established by the present invention;
wherein ,IsFor the source image, ItIn order to be the target image,respectively, the result of each stage transformation.Extracting the network for features, NtranCo-location (POCL) sub-networks for potential objects, TtranFor the corresponding transformation model, NaffiFor Affine Transformation Regression (ATR) sub-networks, TaffiFor the corresponding transformation model, NttpsFor a bidirectional thin-plate spline regression (TTPS) subnetwork, TttpsIs the corresponding transformation model.
FIG. 4 is a schematic diagram of a training triplet and a related transformation model of an OLASA network model in the present invention;
wherein (a) is a source image IsGenerating a reference image Ir(ii) a (b) Are the various transformations between triplets.Respectively representing an affine transformation model and a bi-directional thin-plate spline regression model from a source image to a reference image, as benchmarks (groudtuth) for comparison,respectively representing an affine transformation model and a bidirectional thin plate spline regression model from a source image to a target image,two models representing respective inverse transforms;respectively representing an affine transformation model and a bidirectional thin plate spline regression model from a target image to a reference image,two models of the corresponding inverse transform are represented.
Detailed Description
The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.
The invention provides an image matching method based on a deep semantic alignment network model, which is a deep neural network model OLASA based on semantic alignment, the input of the model method is a source image, a target image and a reference image, OLASA estimates deformation parameters of the source image according to the internal alignment relation of the OLASA through the deep semantic analysis of the source image and the target image, the deformed source image is the output result of the method, and the target object contained in the image can be matched with the corresponding object in the target image. The internal implementation of OLASA is joint learning through three sub-networks: the three sub-networks, potential object co-location (POCL), Affine Transformation Regression (ATR), and bi-directional thin-plate spline regression (TTPS), achieve the purpose of effectively matching images.
Fig. 2 shows a flow of a method for performing image matching on an OLASA network model established by the present invention. The method can be used for any given image pair (namely the source image and the target image), and the image pair formed by the source image and the target image can be obtained by shooting or network downloading and the like. Image data sets such as PF-WILLOW [11], PF-PASCAL (reference [11 ]: Bumsub Ham, Minsu Cho, Cordelia Schmid, and Jean Ponce, "Proposal flow," in CVPR,2016.) and Caltech-101 (reference [12 ]: Li Fei-Fei, Rob Fergus, and Pietro Perona, "One-shot left of object candidates," IEEE TPAMI, 28, No.4,2006.) are used. The reference image is calculated according to the source image in the execution process by adopting the method. The method of the invention is realized by the following steps.
Step 1, extracting semantic features of images
After receiving a source image, a target image and a reference image, OLASA first extracts features of the source image, the target image and the reference image. That is, a pair of images to be matched (source and target images, available as images or shots in an image library)As an input, the method adopts a Convolutional Neural Network (CNN) to extract the characteristics of two images, wherein the network can be the most basic CNN network or the CNN network after improvement or enhancement, which is not general and is called as the CNN network in the textThe two groups of extracted features are namedThese semantic features will be used in the learning process of the subsequent sub-network.
Step 2, estimating the offset by adopting a potential object co-location (POCL) sub-network
In practical situations, objects to be matched in the source image and the target image are often located at different positions in the respective images, that is, the objects to be matched often have a large position difference. The existing method only processes objects to be matched which are almost at the same position, and a special processing method is rarely adopted for the situation with large position difference, so that ideal matching effect is difficult to obtain in practical application of a plurality of methods. The invention aims at the problem that a preprocessing network, a potential object co-location sub-network (POCL), is adopted to eliminate the position deviation of the object to be matched in the first stage of realizing image matching.
The characteristics F of the source image and the target images and FtAs input, estimating potential targets by a classified target detection technique, and training a co-location sub-network based on the potential target locationsAnd then realize Is and ItPreliminary transformation betweenIs estimated.
wherein ,is 4, where (x, y) represents the spatial coordinates of the target image and (x ', y') represents the corresponding sampling coordinates in the source image.
To estimate an offset transformation modelIt is necessary to estimate the position of the object to be matched, for which purpose existing classification target detection techniques can be used, in particular fast R-CNN [12] as a common target detection network]And the like. It should be noted that the target detection technique is not used to obtain an accurate target detection result at this stage, but is used to obtain approximate position information of the potential object to achieve the co-location, so the target detection technique is used differently here. In particular, the amount of the solvent to be used,based on image features Fs and FtOnly two sets of potential object bounding boxes need to be predictedAnd(i=1,…,ns,j=1,…,nt),respectively describe an image Is and ItMiddle (i)th and jthA potential frame, which may be recorded with coordinates of the top left corner and the bottom right corner, ns and ntRespectively represent Is and ItThe number of borders in (1). Again, these borders need not be exact object borders, but rather approximate or possible borders, which in this present stage are intended to estimate their approximate location.
To further locate Is and ItTwo main and semantically related objects in the text, we pass through the latent bounding boxAndthe potential objects are cropped from the image and their size is adjusted to H x W. And then using another feature extraction module, such as CNN,to extract a feature map V corresponding to the source and target imagess i} and {Vt j}. After that, we will describe the child { V }s i} and {Vt jStack (stack) as a feature matrixAndthus, the semantic similarity matrix can be calculated by multiplying the two feature matrices, wherein, focus is on ZstSelecting two corresponding frames as the component items with the highest similarity scoreAndthereby locating two main potential objects separately. Finally, utilizeAndthe spatial coordinates of the position can be calculated to obtain a position deviation transformation modelSource image IsBy passingTransforming into position-shifted images
The POCL is only used to capture the position deviation of the potential object, and when the position deviation is large, the corresponding position adjustment can be realized through the position deviation transformation, but the POCL cannot realize accurate semantic alignment.
Step 3, estimating affine transformation model by utilizing Affine Transformation Regression (ATR) sub-network
OLASA is transformed back by more accurate affineReturn (ATR) subnetworkEstimating affine transformation model of image to be matchedExisting sub-networks [11] that resemble ATR functions]May also be used to estimate the affine transformation. To achieve better results, the method is based on a study of tensor validity [1],[3]An image is formed and ItCharacteristic F ofs 1 and FtForming pairs of features, computing the correlation of these pairs of features, i.e. the 4D correlation tensorEach element of the tensorRecording two local feature vectorsThe inner product of (d) and (d). The eigenvectors and the correlation tensors are L2 normalized. Tensor of degree of correlationIs inputted intoTo estimate and ItAffine transformation model between
Wherein the affine transformation modelIs 6, i.e. 6 affine transformation parameters need to be estimated. Image of a personCan be further transformed by the modelAfter affine transformationAnd ItFurther alignment therewith is accomplished.
Step 4, optimizing alignment effect by using bidirectional thin plate spline regression (TTPS) sub-network
The semantic alignment effect of the image can be further improved or enhanced by utilizing the control point network. In particular, further imaging is performed by using a bi-directional thin-plate spline regression (TTPS) sub-network and ItEstimate a regression transformation model betweenAnd optimizing the image with this model and ItSemantic alignment between them. In fact, existing methods for existing thin-plate spline regression TPS, such as [13 ]]For example, the TPS is unidirectional, the control point is fixed, and the effect is affected accordingly, e.g., excessive distortion occurs in some local areas, resulting in larger object distortion. The present invention contemplates TTPS, also performing regression prediction for a known set of corresponding control points, in combination with a simple example, i.e. on an image and ItA uniform 3 x 3 grid of control points is provided. But unlike TPS, TTPS is bi-directional, adding a slave picture ItTo the imageIn the control point adjustment in the opposite direction, compared with the TPS, the TTPS considers the control points in the two images to be movable, so that the TTPS can effectively remove excessive deformation and improve the matching distortion. In particular implementations, the TTPS is an image and ItIs characterized by and FtThe correlation tensor of the feature pair is the object of computation, namely the 4D correlation tensor To more accurately capture details of the deformation of the object. Using TTPS regression sub-networksEstimating a regression deformation modelCan be represented in form as:
in a similar manner to that described above,is based on and FtCalculating the correlation of the middle control point pair to obtain a regression deformation modelIs obtained by calculating 6 transform weight parameters of the two. According to the estimatedRegression transformation model, canIs further transformed intoThereby realizing the source image and the target image ItIs aligned.
Step 5, joint training of the whole network
The continuous transformation of the three sub-networks realizes a matching principle from coarse to fine, and the image pair Is and ItCan be completely described as the result of three transformation models being connected, i.e. wherein ,representing the composition of the geometric transformation. Moreover, as an integral network, three sub-networks need to be communicated back and forth, and the specific operation method is to perform joint training as an integral network.
Unlike the conventional method that usually uses image pairs as training samples when training the network, for OLASA training, we propose a triple sampling strategy to generate training data by introducing reference images. Each triplet containing a source image IsReference picture IrAnd a target image It. Wherein, from IsBy random geometric transformationOrGenerated reference image IrAs shown in fig. 4 (a). The combination pairs used in the traditional method contain little appearance change, and the triple sampling strategy can simultaneously capture geometric and appearance change in training data, in other words, the triple not only enriches the appearance change of the original image pair, but also provides powerful transition information for the source image and the target image with larger difference. The introduction of triplets enriches the previously proposed transformation model in the specific calculation process, which is separately represented by dashed and solid lines in (b) of fig. 4: 1) the models still to be estimated are, for example, similar to the original affine regression transformation or spline regression transformation, except that the number of models is increased after subdivision, as indicated by the dashed arrowsEtc. 2) since the reference graph is the result of generation, the known transformation can be used as the real labeled data, and the related transformation is represented by the solid arrow, such as
To achieve the training goal of the whole model, we design three loss functions to achieve the optimization of OLASA, including transitivity loss, consistency loss, and alignment loss.
The transmission loss. On the triple sample, we design the transmission loss according to the transmissibility of the geometric transformation and the network MSEBy affine transformationFor example, two other related continuous transforms may be combinedAndto infer a transformation from a source image to a reference imageThat is to say that the first and second electrodes,(x, y) denotes the structure in IsOne point (x, y) in the upper coincidence grid G passes throughShould be equal toAnd aligning projection points after continuous deformation. In which affine transformation regression sub-networksEstimating ternary samplesAndthe grid G (| G | ═ 20 × 20) can be respectively formed byAndand (6) transforming. We use the loss functionI.e., transmission loss, to measure the difference between the two grids before and after the affine transformation.
Similarly, transmission losses will also occur in this processFor spline regressionThe difference analysis of (3).
The consistency is lost. The consistency in the geometric transformation is extended from the cyclic consistency used in the image-to-image transformation, which can be used as a complement to the transmission loss. Loss of consistency of our designIs defined as: given an image Is and ItInter affine transformationAnd corresponding inverse transformationWhen compared to the results of its two-way transformation, the point (x, y) e G should be aligned with the coordinates of the original space, i.e.,thus, the reprojection error of a point in grid G between a source image-target image pair can be calculated by the following formula,
in the ternary sample, other image pairs also contain bidirectional reprojection errors, and the two image pairs are all accumulated to be used as a consistency loss function of the method
Similarly, consistency is also lost in the present methodFor spline regressionThe difference analysis of (3).
The alignment is lost. Loss of alignmentFor evaluatingAndcan further optimize the alignment accuracy.Is based on the document [3]]The Soft-inline Count (Soft-inline Count) and the bidirectional measurement alignment quality are designed, and the calculation formula is as follows:
wherein cstBy aligning IsTo ItSoft inline counting of, similarly, ctsIndicating alignment ItTo IsSoft inline counting of (2).
Step 6, obtaining matching result
Through integral network learning including three sub-networks, the method can obtain an image alignment transformation model according to a training data sample, and can calculate the matching result from a source image to a target image aiming at a given image pair to be matched in a testing stage.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various alternatives and modifications are possible without departing from the invention and scope of the appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (7)
1. An image matching method based on a deep semantic alignment network model comprises the steps of gradually estimating alignment between two semantic similar images by establishing an object position-aware semantic alignment network model OLASA; the method is characterized in that a triple sampling strategy is adopted to train a network model OLASA, and three sub-networks N of potential object co-location POCL, affine transformation regression ATR and bidirectional thin plate spline regression TTPS are adoptedtran,Naffi and NttpsRespectively estimating translation, affine transformation and spline transformation; then, establishing and optimizing an alignment relation between the images in a layering manner to obtain an image matching result; the method comprises the following steps:
step 1, extracting semantic features of an image: a pair of images Is,As convolutional neural networksRespectively extracting two groups of characteristics Fs,Is,ItRespectively a source image and a target image; fs,FtRespectively representing source image semantic features and target image semantic features;
the front end of each sub-network adopts a convolutional neural network CNN to extract the characteristics of the image, and the characteristics are expressed as formula (1):
in formula (1), F is a feature extracted from an image; real numberA data space that is a feature; h. w and d respectively represent the height, width and channel number of the characteristic data space;is a convolutional neural network CNN; i is an image; real numberA data space that is an image; H. w, D respectively represent the height, width, and number of channels of the image data space;
establishing an object location-aware semantic alignment network model OLASA comprising a sub-network Ntran,Naffi and NttpsFor estimating offset, affine and TTPS transformations, respectively, denoted Ttran,Taffi and Tttps(ii) a The source image is transformed to obtain the transformation result of each stageAndfinally by successive transformations THObtaining a source image IsWith the target image ItAlignment result of wherein The method for establishing the object position-aware semantic alignment network model OLASA comprises the following steps of 2-4:
step 2, constructing potential object co-locationSub-network NtranThe system is used for estimating the offset of a target object between images and eliminating the position deviation of an object to be matched;
a source image FsAnd features F of the target imagetCo-locating sub-networks as potential objectsTraining the sub-network based on the potential target object locationAnd then realize Is and ItEstimation of preliminary transformation betweenIs represented as follows:
in the formula ,co-locating a sub-network for the potential object;is Is and ItAn estimate of the preliminary transform in between, i.e. the offset transform; (x, y) represents the spatial coordinates of the target image, (x ', y') represents the corresponding sampling coordinates in the source image;
further location Is and ItTwo semantically related objects in the database, and extracting corresponding feature descriptors { V }s i} and {Vt jWherein i, j represent in the source image and the target image respectivelyThe serial number of the feature point of (1);
will describe the son { Vs i} and {Vt jStack as feature matrixAndcalculating a semantic similarity matrix by multiplying two feature matrices:
selection of ZstForming a similar feature pair group by a plurality of feature pairs with the highest similarity scores, calculating two corresponding regions by using feature point groups respectively represented in a source image and a target image, taking the two corresponding regions as two main potential objects, and calculating corresponding frame coordinates, namely space positions;
and then calculating a position offset transformation model by using the space coordinates of the corresponding frame of the potential object obtained by positioningSource image IsBy passingTransforming into position-shifted images
Step 3, constructing affine transformation regression sub-network NaffiEstimating affine transformation model of image to be matched using affine transformation regression subnetworkObtaining affine transformation parameter estimation; the method comprises the following steps:
affine transformation regression sub-network for image features and FtConstructing feature pairs, calculating the correlation of the feature pairsEstimating parameters of the affine transformation model;
degree of correlation of feature pairsI.e. the 4D correlation tensor,each element of the tensorRecording two local feature vectors Inner product of (2);
the eigenvectors and the correlation tensors are L2 normalized; tensor of degree of correlationInput deviceAccording to formula (3) and ItInter affine transformation model estimation
Image of a personFurther transformed by the formula (3)After affine transformationFurther with ItAligning;
step 4, constructing a bidirectional thin plate spline regression subnetwork Nttps;NttpsEstimating a regression transformation model between the affine transformed image and the target image by using the control points, and further optimizing the semantic alignment effect of the image;
in particular in images and ItA control point grid is arranged on the upper part so as to and FtCorrelation tensor of pairs of featuresFor calculating objects, adoptEstimating a regression deformation modelIs represented by formula (4):
wherein ,according to and FtCalculating the relevance of the middle control point pair; obtaining a regression deformation model by calculating 6 transformation weight parametersThe estimation result of (2); according to regression transformation modelWill be provided withIs further transformed intoRealizing source image and target image ItAlignment of (2);
step 5, introducing a reference image, selecting a training sample by adopting a triple method, and jointly training a network model OLASA;
step 51, generating training data by adopting a triple sampling strategy method; each triplet containing a source image IsReference picture IrAnd a target image It(ii) a From IsBy random geometric transformationOrGenerating a reference image Ir;
Step 52, designing three loss functions to realize the optimization of OLASA, including the transitive loss functionNumber ofConsistency loss functionAnd alignment loss function
Transfer loss functionFor measuring the accuracy of the predicted transform model; taking the difference between the transformation result of the transformation model and the transformation result of another transformation path for achieving the same transformation purpose as the transmission loss; another transformation approach specifically employs an indirect transformation approach through a combination of two transformation models; the indirect conversion way realizes the same conversion purpose as the direct conversion way which only adopts one conversion model, namely the conversion target of the conversion model to be detected is the same;
consistency loss functionIs defined as: in a ternary sample consisting of a source image, a target image and a reference image, the accumulation of the bidirectional reprojection error of a point in the grid G between different image pairs constitutes a consistency loss functionFor representing affine transformationsSum spline regressionA difference of (a);
alignment loss functionAlignment quality design based on soft inline counting and two-way measurement for evaluationAndthe alignment accuracy is further optimized;calculated by equation (8):
wherein ,cstBy aligning IsTo ItSoft inline counting of (2); c. CtsIndicating alignment ItTo IsSoft inline counting of (2);
step 6, aiming at a given image pair to be matched, obtaining an alignment result from a source image to a target image by utilizing the semantic alignment network model OLASA of the object position perception trained in the step 5, namely obtaining an image matching result;
through the steps, the image matching based on the deep semantic alignment network model is realized.
2. The image matching method based on the deep semantic alignment network model as claimed in claim 1, wherein in step 2, the position of the potential target object is obtained by estimating the potential target through a classified target detection method.
3. The image matching method based on the deep semantic alignment network model as claimed in claim 1, wherein in step 2, I is further positioneds and ItThe two semantically related objects are extracted by a feature extraction module CNN,extracting to obtain corresponding feature mapping { Vs i} and {Vt j}。
5. The image matching method based on the deep semantic alignment network model as claimed in claim 1, wherein in step 4, the bidirectional thin plate spline regression sub-network adopts a bidirectional strategy to increase the slave image ItTo the imageAnd the control point adjustment in the opposite direction effectively avoids the excessive distortion or distortion of the image.
6. The image matching method based on deep semantic alignment network model as claimed in claim 1, wherein in step 52, for affine transformationImage IsOne point in a grid G composed of the pixel points in (1) is expressed by (x, y), and calculation is performed based on the grid pointBy passingAndjudgment ofAccuracy, i.e. aboutAnd as a function of error loss, transfer lossRepresented by formula (5):
wherein ,an affine transformation model of a source image and a target image;an affine transformation model for the target image and the reference image;obtaining an affine transformation model from a source image to a target image according to the labeling data;
7. As claimed in claim 6The image matching method based on the deep semantic alignment network model is characterized in that in step 52, a consistency loss functionRepresented by formula (7):
wherein ,ε(Is,It) The reprojection error for a point in the grid G between the source image-target image pair can be calculated by equation (6):
wherein ,ε(Is,It) Is the reprojection error between the source image-target image pair;as an image Is and ItAffine transformation between;is composed ofCorresponding inverse transformation;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110516741.5A CN113313147B (en) | 2021-05-12 | 2021-05-12 | Image matching method based on depth semantic alignment network model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110516741.5A CN113313147B (en) | 2021-05-12 | 2021-05-12 | Image matching method based on depth semantic alignment network model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113313147A true CN113313147A (en) | 2021-08-27 |
CN113313147B CN113313147B (en) | 2023-10-20 |
Family
ID=77373055
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110516741.5A Active CN113313147B (en) | 2021-05-12 | 2021-05-12 | Image matching method based on depth semantic alignment network model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113313147B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111621508A (en) * | 2020-06-11 | 2020-09-04 | 云南中烟工业有限责任公司 | Tobacco terpene synthase NtTPS7 gene and vector and application thereof |
CN115861393A (en) * | 2023-02-16 | 2023-03-28 | 中国科学技术大学 | Image matching method, spacecraft landing point positioning method and related device |
CN116977652A (en) * | 2023-09-22 | 2023-10-31 | 之江实验室 | Workpiece surface morphology generation method and device based on multi-mode image generation |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107862707A (en) * | 2017-11-06 | 2018-03-30 | 深圳市唯特视科技有限公司 | A kind of method for registering images based on Lucas card Nader's image alignment |
US20190371080A1 (en) * | 2018-06-05 | 2019-12-05 | Cristian SMINCHISESCU | Image processing method, system and device |
CN110580715A (en) * | 2019-08-06 | 2019-12-17 | 武汉大学 | Image alignment method based on illumination constraint and grid deformation |
CN110909778A (en) * | 2019-11-12 | 2020-03-24 | 北京航空航天大学 | Image semantic feature matching method based on geometric consistency |
CN112102303A (en) * | 2020-09-22 | 2020-12-18 | 中国科学技术大学 | Semantic image analogy method for generating countermeasure network based on single image |
CN112634341A (en) * | 2020-12-24 | 2021-04-09 | 湖北工业大学 | Method for constructing depth estimation model of multi-vision task cooperation |
-
2021
- 2021-05-12 CN CN202110516741.5A patent/CN113313147B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107862707A (en) * | 2017-11-06 | 2018-03-30 | 深圳市唯特视科技有限公司 | A kind of method for registering images based on Lucas card Nader's image alignment |
US20190371080A1 (en) * | 2018-06-05 | 2019-12-05 | Cristian SMINCHISESCU | Image processing method, system and device |
CN110580715A (en) * | 2019-08-06 | 2019-12-17 | 武汉大学 | Image alignment method based on illumination constraint and grid deformation |
CN110909778A (en) * | 2019-11-12 | 2020-03-24 | 北京航空航天大学 | Image semantic feature matching method based on geometric consistency |
CN112102303A (en) * | 2020-09-22 | 2020-12-18 | 中国科学技术大学 | Semantic image analogy method for generating countermeasure network based on single image |
CN112634341A (en) * | 2020-12-24 | 2021-04-09 | 湖北工业大学 | Method for constructing depth estimation model of multi-vision task cooperation |
Non-Patent Citations (2)
Title |
---|
刘岩;吕肖庆;秦叶阳;汤帜;徐剑波;: "尺度与颜色不变性图像特征描述", 小型微型计算机系统, no. 10, pages 187 - 192 * |
廖明哲;吴谨;朱磊;: "基于ResNet和RF-Net的遥感影像匹配", 液晶与显示, no. 09, pages 91 - 99 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111621508A (en) * | 2020-06-11 | 2020-09-04 | 云南中烟工业有限责任公司 | Tobacco terpene synthase NtTPS7 gene and vector and application thereof |
CN111621508B (en) * | 2020-06-11 | 2022-07-01 | 云南中烟工业有限责任公司 | Tobacco terpene synthase NtTPS7 gene and vector and application thereof |
CN115861393A (en) * | 2023-02-16 | 2023-03-28 | 中国科学技术大学 | Image matching method, spacecraft landing point positioning method and related device |
CN115861393B (en) * | 2023-02-16 | 2023-06-16 | 中国科学技术大学 | Image matching method, spacecraft landing point positioning method and related device |
CN116977652A (en) * | 2023-09-22 | 2023-10-31 | 之江实验室 | Workpiece surface morphology generation method and device based on multi-mode image generation |
CN116977652B (en) * | 2023-09-22 | 2023-12-22 | 之江实验室 | Workpiece surface morphology generation method and device based on multi-mode image generation |
Also Published As
Publication number | Publication date |
---|---|
CN113313147B (en) | 2023-10-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11763433B2 (en) | Depth image generation method and device | |
Labbé et al. | Cosypose: Consistent multi-view multi-object 6d pose estimation | |
CN113313147B (en) | Image matching method based on depth semantic alignment network model | |
CN111625667A (en) | Three-dimensional model cross-domain retrieval method and system based on complex background image | |
CN113160285B (en) | Point cloud matching method based on local depth image criticality | |
CN103700099A (en) | Rotation and dimension unchanged wide baseline stereo matching method | |
CN107766864B (en) | Method and device for extracting features and method and device for object recognition | |
Yi et al. | Motion keypoint trajectory and covariance descriptor for human action recognition | |
CN104517289A (en) | Indoor scene positioning method based on hybrid camera | |
CN113361542A (en) | Local feature extraction method based on deep learning | |
CN110969648A (en) | 3D target tracking method and system based on point cloud sequence data | |
CN110544202A (en) | parallax image splicing method and system based on template matching and feature clustering | |
CN111368733B (en) | Three-dimensional hand posture estimation method based on label distribution learning, storage medium and terminal | |
CN116342648A (en) | Twin network target tracking method based on mixed structure attention guidance | |
Shen et al. | Semi-dense feature matching with transformers and its applications in multiple-view geometry | |
Lee et al. | Learning to distill convolutional features into compact local descriptors | |
CN112329662B (en) | Multi-view saliency estimation method based on unsupervised learning | |
CN113988269A (en) | Loop detection and optimization method based on improved twin network | |
Huang et al. | Life: Lighting invariant flow estimation | |
Zhang et al. | An automatic three-dimensional scene reconstruction system using crowdsourced Geo-tagged videos | |
CN110849380A (en) | Map alignment method and system based on collaborative VSLAM | |
Xu et al. | Local feature matching using deep learning: A survey | |
Xu et al. | Improved HardNet and Stricter Outlier Filtering to Guide Reliable Matching. | |
CN115375746A (en) | Stereo matching method based on double-space pooling pyramid | |
CN114155406A (en) | Pose estimation method based on region-level feature fusion |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |